基于孪生网络的人工震源分布式光纤传感数据噪声压制  

Siamese network based noise elimination of artificial seismic data recorded by distributed fiber-optic acoustic sensing

作　　者：邵婕 王一博[1,2] 梁兴[3] 薛清峰[1,2] 梁恩茂 史树有 SHAO Jie;WANG YiBo;LIANG Xing;XUE QingFeng;LIANG EnMao;SHI ShuYou(Institute of Geology and Geophysics,Chinese Academy of Sciences,Beijing 100029,China;Innovation Academy for Earth Science,Chinese Academy of Sciences,Beijing 100029,China;Zhejiang Oilfield Branch,China National Petroleum Corporation,Hangzhou 311100,China;Beijing Global Safety Technology Co.,Ltd.,Beijing 100094,China)

机构地区：[1]中国科学院地质与地球物理研究所,北京100029 [2]中国科学院地球科学研究院,北京100029 [3]中国石油天然气股份有限公司浙江油田分公司,杭州311100 [4]北京辰安科技股份有限公司,北京100094

出　　处：《地球物理学报》2022年第9期3599-3609,共11页Chinese Journal of Geophysics

基　　金：国家重点研发计划(2021YFA0716800);中国科学院稳定支持基础研究领域青年团队计划(YSBR-020);中国石油天然气集团有限公司重大工程技术专项“井下光纤智能化监测技术现场试验”(2020F-44)资助。

摘　　要：受耦合效应和其他诸多因素的影响,野外实际采集的分布式光纤声波传感(Distributed Fiber-optic Acoustic Sensing,DAS)数据的信噪比通常较低.因此,在后续处理之前,需要首先对DAS数据进行去噪处理.传统的基于人工智能监督学习的去噪方法能够对DAS数据中的噪声进行压制,但它需要大量含噪声和无噪声数据进行成组标记,人工标记工作量巨大.为此,本文提出了基于自监督学习孪生网络的DAS地震数据去噪方法.该方法基于自监督学习中的孪生网络结构,采用U-Net网络建立深度学习框架.所提框架通过对输入的含噪声数据进行数据自我特征学习,迭代获取去噪目标函数的最优解,从而实现自监督高精度深度学习去噪网络构建.合成数据和实际资料处理结果表明,本文方法可以有效抑制人工震源DAS采集数据中的随机噪声,显著提高去噪结果的信噪比和同相轴的连续性.此外,本文方法能够避免常规监督学习方法需要进行数据标记的人工工作量,有效提高实际地震数据去噪处理效率.The signal-to-noise ratio(SNR)of field distributed fiber-optic acoustic sensing(DAS)data is usually relatively low due to the poor coupling and many other factors.Therefore,it is necessary to denoise the field DAS seismic data before the subsequent processing.The traditional artificial intelligence supervised learning-based denoising method can effectively eliminate noise in the DAS data.However,it needs to label the data using large groups of noisy and noise-free data,which turns out to be a huge effort.This paper proposes a noise suppression method for DAS seismic data using a self-supervised siamese network.Based on the theory of siamese network in self-supervised learning,the network structure is built using a U-Net architecture.The optimal solution of the denoising objective function is obtained iteratively by learning characteristics of the input noisy data from itself.Then a self-supervised high-precision deep learning-based denoising network is established.Results of synthetic data and field data show that the proposed method can effectively suppress the random noise of artificial seismic data recorded by DAS,and improve the SNR and continuity of seismic events.In addition,the proposed method avoids manual labeling in the traditional supervised learning methods and significantly improves the efficiency of seismic data denoising.

关 键 词：孪生网络 自监督深度学习 分布式光纤声波传感 随机噪声压制 

分 类 号：P631[天文地球—地质矿产勘探]